Tidewave MCP Integration

Runtime intelligence for Phoenix apps via MCP. Prefer Tidewave tools over Bash when available.

Iron Laws — Never Violate These

1. DEV ONLY — Never use Tidewave tools in production contexts. Avoid on shared dev servers with production data copies
2. PREFER TIDEWAVE OVER BASH — mcp__tidewave__get_docs > web_fetch, execute_sql_query > psql
3. CHECK AVAILABILITY FIRST — Use /mcp command or detect mcp__tidewave__ tools
4. SQL IS READ-HEAVY — Use execute_sql_query for SELECT, be careful with mutations
5. EXACT VERSIONS — get_docs returns docs for YOUR mix.lock versions, not latest

Quick Reference

Detection

# Check endpoint
curl -s http://localhost:4000/tidewave/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"ping"}'

Or use /mcp in Claude Code to see connected servers.

Essential Patterns

Test Function Immediately

# mcp__tidewave__project_eval
MyApp.Accounts.create_user(%{email: "[email protected]"})

Verify Migration

-- mcp__tidewave__execute_sql_query
SELECT column_name, data_type FROM information_schema.columns
WHERE table_name = 'users';

Debug LiveView (with PID from browser)

# mcp__tidewave__project_eval
pid = pid("0.1234.0")
:sys.get_state(pid) |> Map.get(:socket) |> Map.get(:assigns) |> Map.keys()

Setup Requirements

# mix.exs
{:tidewave, "~> 0.1", only: :dev}

# endpoint.ex (in dev block)
plug Tidewave

# config/dev.exs (for LiveView source mapping)
config :phoenix_live_view,
  debug_heex_annotations: true,
  debug_attributes: true

Reliability Guards

Worktree/port check (FIRST, in multi-worktree setups): multiple worktrees = multiple dev servers on different ports. Before trusting any Tidewave result, confirm the endpoint belongs to THIS checkout: grep config/dev.exs for the configured port, and verify with project_eval File.cwd!() — if it returns a different worktree path, you're debugging the wrong server.

Schema introspection BEFORE SQL: never guess column names. Run get_ecto_schemas (or query information_schema.columns) before writing SQL against a table you haven't already introspected this session. A guessed-column error costs more than the introspection.

Output-size guard: runtime output is unbounded. Always cap it — LIMIT 20 in SQL, Enum.take(20) in evals, inspect(x, limit: 50, printable_limit: 500) for large structs. Re-query narrower rather than dumping wide.

browser_eval fallback: if mcp__Tidewave-Web__browser_eval is absent or errors, don't stall — inspect the same state server-side: LiveView assigns via :sys.get_state(pid) in project_eval, rendered HTML via Phoenix.LiveViewTest, or read the template source directly.

QA walkthrough pattern: after a feature completes, run a short checklist through project_eval/browser_eval: create the record, fetch it back, exercise the main event, check get_logs level: :error is clean. Report each step's pass/fail — not just "smoke test passed".

Proactive Runtime Checks

Don't just use Tidewave reactively. Query runtime state at workflow checkpoints automatically:

• After code edits: get_logs level: :error (catch runtime crashes)
• After features complete: project_eval smoke test (behavioral check)
• Before planning: get_ecto_schemas + routes eval (concrete context)
• When investigating: Auto-capture errors before asking user
• LiveView UI bugs: browser_eval to inspect DOM state before editing components

See ${CLAUDE_SKILL_DIR}/references/proactive-patterns.md for full integration points.

References

For detailed patterns, see:

• ${CLAUDE_SKILL_DIR}/references/proactive-patterns.md - Push-like runtime patterns at workflow checkpoints
• ${CLAUDE_SKILL_DIR}/references/tool-examples.md - Complete tool usage examples
• ${CLAUDE_SKILL_DIR}/references/validation-checklist.md - Runtime validation patterns